1 / 37

Higher-Order Tree Transducers and Their Expressive Power

Higher-Order Tree Transducers and Their Expressive Power. Kazuhiro Inaba at Dagstuhl Seminar, May 2013. Regular Tree Grammar. (Top-Down) Tree Transducer. input tree. Δ = { a , b , c , ...} N = {F, G, H, ..., S, T, ...} R = set of rules S ∈ N. Σ = { s , t , u , ...}

conor
Download Presentation

Higher-Order Tree Transducers and Their Expressive Power

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Higher-Order Tree TransducersandTheir ExpressivePower Kazuhiro Inaba at Dagstuhl Seminar, May 2013

  2. RegularTree Grammar (Top-Down)Tree Transducer input tree • Δ = {a, b, c, ...} • N = {F, G, H, ..., S, T, ...} • R = set of rules • S ∈ N • Σ = {s, t, u, ...} • Δ = {a, b, c, ...} • N = {F, G, H, ..., S, T, ...} • R = set of rules • S ∈ N a S  a T T S  b T  c S S • S (sx1x2)  a(T x1) (T x2) • S (t)  b • T (sx1x2)  c(S x1) (S x2) c c b b a b Each nonterminal generates(a set of) trees. F : O Each nonterminal takes an input tree and generates(a set of) trees. F : I O

  3. RegularTree Grammar (Top-Down)Tree Transducer input tree tree parameter tree parameter MacroTree Transducer Context-FreeTree Grammar input tree S  Tdd T y1 y2  T (b y1) (c y2) T y1 y2  a y1 y2 • S (s x1)  T x1dd • T (s x1) y1 y2 T x1(b y1) (c y2) • T z y1y2 a y1 y2 Each nonterminal takes parameter trees and generates (a set of) trees. F : Ok O Each nonterminal takes an input tree and parameter trees, and generates trees. F : IOk O

  4. RegularTree Grammar (Top-Down)Tree Transducer input tree tree parameter tree parameter MacroTree Transducer Context-FreeTree Grammar input tree higher-order parameter higher-order parameter multipleinput trees Higher-OrderTree Transducer Higher-OrderTree Grammar input tree HMTT

  5. Example of a higher-order transducer pair Mult: I O Mult (pairx1x2) Iter x1 (Add x2) z Iter: I (O O)  O O Iter (sx1) f y Iter x1 f (f y) Iterz f y y Add : I O O Add (sx1) y  Add x1 (sy) Add z y  y S S S S S S S Z S Z S S S Z

  6. (Examples of) problems we should be interested in Membership Given a higher-order grammar G and a tree t, decide whether t ∈ [G] or not. Type Checking Given a higher-order tree transducer f and regular tree languages S and T, decide whether f(S)⊆T or not. Model Checking Given a deterministic higher-order grammar Grepresenting a (possibly infinite) single tree t, and a MSO sentence φ, decide whether t satisfies φ. Equi-Expressivity What is the automata-like of the models? Can they be “decomposed” to simpler models?

  7. Agenda • Introduction • Two notions of “higher-order” types. • Review of known results. • Context-sensitiveness of “safe” higher-order languages [I. and Maneth, 2008]

  8. Two Notions of “Higher Order” Types (1) • D0 = O • “Trees” are order-0. • Di+1= { Dik Di | k ∈ N } • Functions from order-i objects to order-i objects are order-(i+1). • order(t) = i if t ∈ Di • “Derived Types” • OI-Hierarchy [Damm 82] • High-Level Tree Tranducer [Engelfriet & Vogler 88]

  9. Two Notions of “Higher Order” Types (2) • Recently actively studied in context of program verification [Ong 06, ...] or linguistics [Kobele&Salvati 13, ...]. • D ::= O | DD • “Trees” are order-0, and ... • order(O) = 0 • order(t1t2) = max(order(t1)+1, order(t2))

  10. The Difference • Functions parameters of “Derived Types” have decreasing order Dn (Dn-1  (Dn-2  ... (O  O) ...)) which does not contain, e.g.,λx. λf. λy. f x : O (OO)  O  O It implies: Safety [Knapik&Niwinski&Urzyczyn01, 02] No order-k subterm can contain order <k free variables.

  11. Safety Safety [KNU 01, 02] No order-k subterm can contain order <k free variables. Unsafe example: λy. ((λx. λy. a x y) y) • λy. ( (λy. a x y)[x/y] ) • λy. ( λy. a y y ) this is wrong [KNU 01, 02] [Blum&Ong 09] In safe grammars/λ-calculus, you don’t need to care about variable capturing while substitution. !!

  12. “Safe” :: Di+1 = {Dik Di} “Unsafe” :: DD MSO model checking is decidable. [Ong 06, Kobayashi 09] Hierarchy is strict. [Kartzow&Parys 12] Equivalent to “collapsible pushdown automata” [Hague&Murawski&Ong&Serre 08] ???? ???? Grammars • MSO model checking is decidable. [KNU 01, 02] • Hierarchy is strict. [Damm 82] • Equivalent to “iterated pushdown automata” [Da 82] (= (stack of)* stacks) • Context-sensitive.[Maneth 02][I.&Maneth 08] Transducers [EV88] • n-DHTT = (1-DHTT) n • n-NHTT ⊆ (1-NHTT) n

  13. “Safe” :: Di+1 = {Dik Di} “Unsafe” :: DD MSO model checking is decidable. [Ong 06, Kobayashi 09] Hierarchy is strict. [Kartzow&Parys 12] Equivalent to “collapsible pushdown automata” [Hague&Murawski&Ong&Serre 08] ???? ???? Grammars • MSO model checking is decidable. [KNU 01, 02] • Hierarchy is strict. [Damm 82] • Equivalent to “iterated pushdown automata” [Da 82] (= (stack of)* stacks) • Context-sensitive.[Maneth 02][I.&Maneth 08] Transducers [EV88] • n-DHTT = (1-DHTT) n • n-NHTT ⊆ (1-NHTT) n

  14. “Safe” :: Di+1 = {Dik Di} “Unsafe” :: DD MSO model checking is decidable. [Ong 06, Kobayashi 09] Hierarchy is strict. [Kartzow&Parys 12] Equivalent to “collapsible pushdown automata” [Hague&Murawski&Ong&Serre 08] ???? ???? Grammars • MSO model checking is decidable. [KNU 01, 02] • Hierarchy is strict. [Damm 82] • Equivalent to “iterated pushdown automata” [Da 82] (= (stack of)* stacks) • Context-sensitive.[Maneth 02][I.&Maneth 08] Transducers [EV88] • n-DHTT = (1-DHTT) n • n-NHTT ⊆ (1-NHTT) n

  15. “Collapsible” Pushdown Automata [Hague at al. 08] • Order-n collapsible pushdown store is • (stack of)n symbols • with each symbol associated with “links” • Push1: pushes a symbol and link to the top. • Dupk: duplicates the top order-k stack. • Popk: pops the top order-k stack. • Collapse : moves the top to the pointee of the top link.

  16. “Safe” :: Di+1 = {Dik Di} “Unsafe” :: DD MSO model checking is decidable. [Ong 06, Kobayashi 09] Hierarchy is strict. [Kartzow&Parys 12] Equivalent to “collapsible pushdown automata” [Hague&Murawski&Ong&Serre 08] ???? ???? Grammars • MSO model checking is decidable. [KNU 01, 02] • Hierarchy is strict. [Damm 82] • Equivalent to “iterated pushdown automata” [Da 82] (= (stack of)* stacks) • Context-sensitive.[Maneth 02][I.&Maneth 08] Transducers [EV88] • n-DHTT = (1-DHTT) n • n-NHTT ⊆ (1-NHTT) n 2-unsafe = 2-safe [Aehlig&Miranda&Ong 05] 2-unsafe-det ⊄ n-safe-det[Parys 11, 12]

  17. “Safe” :: Di+1 = {Dik Di} “Unsafe” :: DD MSO model checking is decidable. [Ong 06, Kobayashi 09] Hierarchy is strict. [Kartzow&Parys 12] Equivalent to “collapsible pushdown automata” [Hague&Murawski&Ong&Serre 08] ???? ???? Grammars • MSO model checking is decidable. [KNU 01, 02] • Hierarchy is strict. [Damm 82] • Equivalent to “iterated pushdown automata” [Da 82] (= (stack of)* stacks) • Context-sensitive.[I.&Maneth 08] Transducers [EV88] • n-DHTT = (1-DHTT) n • n-NHTT ⊆ (1-NHTT) n

  18. 1st order Decomposition of Safe HTT [Engelfrier&Vogler 86,88] [Caucal 02] Safe-n-DHTT =(Safe-1-DHTT)n Safe-n-NHTT⊆ (Safe-1-NHTT)n n-thorder tree transducer is representable by a n-fold composition of 1st-order tree transducers. Note: Higher order grammars can be simulated by Out(HTT).

  19. Proof: n-HTT = (n-1)-HTT ; 1-HTT Idea:Represent 1st-order term TreeTree by a Tree. Represent 1st-order application symbolically, too. F :: Tree Tree F zs (s y) F :: Tree TreeTree Fz y s (s y) …  @ (F x) z …  F x z

  20. Proof: n-HTT = (n-1)-HTT ; 1-HTT Represent 1st-order things symbolically. Then a 1-HTT performs the actual “application”. F :: Tree Tree F zs (s y) …  @ (F x)z Eval(@ f b) y Eval f (Evalb y) Evaly y  y Eval(sx) y s (Evalx y) Evalz y  z

  21. Mult(pair x1x2)@ (Iter x1(Add x2))z Iter(s x) f@ (Iter x f)(@ f y) Iterz f y Add (s x)@ (Add x) (s y) Add zy Example @ Mult(Pair (s z) (s z)) Z @ @ @ Iter(s z) (Add (s z)) @ @ @ Z Z Z @ @ @ Iterz (Add (s z)) @ y y y Add (s z) Add (s z) y y y S y

  22. Eval(@ f b) y  Eval f (Eval b y) Evaly y  y Eval(sx) y  s (Eval x y) Evalzy  z S Z Eval( , y=⊥) @ Eval( , y= ) @ @ Z Z Eval( ,y=Eval( ,y= ) Eval( ,y= ) @ @ @ @ Z Z @ @ @ @ y y y y y y y y y y y S S S S y y y y

  23. If it’s unsafe... @ S  F a F y  b (H (G y) c) G x y ... H f y  d (f e) y b a @ H (G y) c F a @ S  @ F a F  b (@ (H (@ G y)) c) G  .. H f  d (@ f e) y a b @ d ((G y) e) y c

  24. Consequences of First-Order Decomposition • Proof: because MTT (1-HTT) has the property. • This gives decidable “type checking”. • f(S)⊆T  f-1(T)⊆S : inclusion of REG is decidable. [EV88] f-1(T) ∈ REG if f∈Safe-HTT and T∈REG

  25. Consequences of First-Order Decomposition • DCS = Deterministic-Context-Sensitive= DLINSPACE membership • Proof: in a next few slides... • Corollary : Safe Higher-order languages (aka. OI-Hierarchy) are context-sensitive. [I. and Maneth 08] f(T) ∈ DCS if f∈Safe-HTT and T∈REG

  26. Proof: Out(1-HTT n) ∈DLINSPACEThe Key Idea n-HTT λλ λ s0 1-HTT n “garbage free” 1-HTT n τ1 τ2 τn transform t t s s s2 Sn-1 s1 τ'del τ'1 τ'2 τ'n

  27. How to Construct the “Garbage-Free” Form Make each 1-HTT “productive” τn-1 τn t τn-1 τ’n t

  28. How to Construct the “Garbage-Free” Form Make each 1-HTT “productive”by separating its “deleting” part = τn τ’del τ’n τn-1 τn t τn-1 τ’del τ’n t

  29. How to Construct the “Garbage-Free” Form Make each 1-HTT “productive”by separating its “deleting” part,and fuse the deleter to the left [En75,77][EnVo85][EnMa02] τn-1 τn t τ’n-1+del τ’n t

  30. Repeat τ1 τ2 τ3 τ4 Split τ’4d τ1 τ2 τ3 τ’4 Fuse τ1 τ2 τ34d τ’4 Split τ’34d τ1 τ2 τ’3 τ’4 Fuse τ1 τ234d τ’3 τ’4 Split τ’234d τ1 τ’2 τ’3 τ’4 Fuse τ1234d τ’2 τ’3 τ’4 Split τ’1234d τ’1 τ’2 τ’3 τ’4

  31. Summary: Out(Safe-n-HTT) is context sensitive. λ λ λ • Decompose n-HTT to (1-HTT)n. • Split each 1-HTT to (LT; 1-HTT). = deleting and productive part • Fuse deleting part ahead. 1-HTT ; LT ⊆ 1-HTT • Now all intermediatetrees must be small.Try them all in DLINSPACE. τ1 τ2 τn τ'del τ'1 τ'2 τ'n t s2 Sn-1 s1 s0

  32. Open Questions “Safe” :: Di+1 = {Dik Di} “Unsafe” :: DD MSO model checking is decidable. [Ong 06, Kobayashi 09] Hierarchy is strict. [Kartzow&Parys 12] Equivalent to “collapsible pushdown automata” [Hague&Murawski&Ong&Serre 08] Is unsafe higher-order languages context-sensitive? Does unsafe higher-order transducers have first-order decompositions? Grammars • MSO model checking is decidable. [KNU 01, 02] • Hierarchy is strict. [Damm 82] • Equivalent to “iterated pushdown automata” [Da 82] (= (stack of)* stacks) • Context-sensitive.[Maneth 02][I.&Maneth 08] Transducers [EV88] • n-DHTT = (1-DHTT) n • n-NHTT ⊆ (1-NHTT) n

  33. Idea: Stack-TT • 1-HTT has difficulty in implementing capture-avoiding substitution. • How about extending them with a stack • f (a x1 ... xn) y1 ... ymys RHS where RHS ::= d RHS ... RHS | yi | f xi RHS ... RHS ys POP m values PUSH

  34. Unsafe substitution De-Bruijn index + Stack-TT (λx. λy. (A x y)) BC A B C Eval s s (λ. λ. (A 0 1)) BC z @ Eval(@ f a) ys  Evalf (Eval a ys) ys Eval(s x) y ys  Evalx ys Evalz y ys  y @ C PUSH! A B s s POP! z

  35. Even if it’s unsafe... S  @ F a F  b (@ (H (@ G 2)) c) G  .. H f  d (@ (k1d2 f) e) 0 S  F a F y  b (H (G y) c) G x y ... H f y  d (f e) y Eval (k1d2 x) y1 y2 y3 ys = Eval x y1 ys @ S  @ F a F  b (@ (H (@ G y)) c) G  .. H f  d (@ f e) y a b @ d ((G y) e) y c

  36. Found an error in the proof during discussion in the seminar... X( Pros & Cons Theorem: Unsafe-n-HTT ⊆ (Stack-TT)n • Good: • Good: • Bad: it is hard to make it garbage-free. • There can be a sequence of significant stack operations not generating output. • Bad: it may be overly powerful. Theorem: (Stack-TT ; LT) ⊆ Stack-TT Theorem: Stack-TT ⊄Unsafe-n-HTT Proof:inverse of stack-TT does not preserve regularity, while unsafe-n-HTT does.

  37. Summary • “Safe” and “unsafe” HTT are different. • Can we transfer results in “safe” case to “unsafe”? • Can we decompose an unsafe-HTT to 1st order machines? • Can we show context-sensitivity of the output language? • See http://www.kmonos.net/pub/tmp/smtt.pdf for the technical development.

More Related